MRI for Prostate Cancer: Background, Current Uses, and Integration With AI

The utilization and popularity of multiparametric magnetic resonance imaging (mpMRI) of the prostate has grown substantially in the last decade. An estimated 280,000 new cases of prostate cancer will be diagnosed in 2023, and the incorporation of mpMRI into the clinical pathway remains a critical component for many clinicians and patients across the United States.¹ In fact, the latest American Urological Association (AUA)/Society of Urologic Oncology (SUO) guideline on early detection advises clinicians to utilize MRI prior to initial prostate biopsy to increase the detection of clinically significant (Gleason grade group [GG] 2 or higher) prostate cancer.²

mpMRI refers to a technique that combines a variety of MRI imaging modalities, providing a comprehensive, noninvasive image of the prostate. Specifically, sequences such as T2-weighted imaging, diffusion-weighted imaging (DWI), and dynamic contrast-enhanced imaging offer specific contributions to help increase diagnostic accuracy. For example, T2-weighted imaging is utilized to identify prostatic zonal anatomy. In the peripheral zone tumor setting, signal intensity is normally high in normal tissue, but would appear as low in cancer suspicious lesions.³

DWI is another example. The premise of DWI is that restricted diffusion of water is more prominent in tissues with high cell densities, such as tumors.³ DWI facilitates the delineation of normal prostatic tissue from potentially cancerous tissue.³

With regard to prostate cancer diagnosis, mpMRI helps facilitate the identification of clinically significant prostate cancer that traditional methods, such as transrectal ultrasound-guided biopsy, may have missed due to the stochastic nature of systematic sampling. The combination of the 2 techniques can help identify and target suspicious areas within the prostate. Those areas are then reported using the uniform Prostate Imaging-Reporting and Data System (PI-RADS)—the recommended reporting scheme for mpMRI according to the AUA/SUO guideline²—to help identify clinically significant prostate cancer.

Also according to the AUA/SUO guideline, active surveillance is the preferred management strategy for low-risk prostate cancer. It is an option for some patients with intermediate-risk prostate cancer as well.⁴ Active surveillance facilitates the close monitoring of a patient’s disease without upfront treatment, allowing the patient to live free of the complications associated with common prostate cancer treatments such as prostatectomy or prostate radiation therapy.

Recently, there has been significant excitement in the news regarding artificial intelligence (AI) and its role in medicine. AI has been front and center at the AUA Annual Meeting, with several presenters showcasing their latest and greatest research using AI in the clinical setting in urology every year. Recently, Dr. Hung and colleagues presented data on utilizing AI to help analyze automated performance metrics during robotic surgery. Specifically, the authors demonstrated that AI can be used to analyze surgeon performance during robot-assisted radical prostatectomy (RALP) by taking automated performance metrics (APMs) from the robotic platform. In conjunction with clinical features, AI can predict the return of urinary continence at 3 and 6 months.⁵ Furthermore, a follow-up publication on AI by the same authors utilized the same concept of combining AI with APMs to facilitate training surgeons to improve their vesicourethral anastomotic technique during RALP.⁶

What role can AI play in the diagnosis and management of prostate cancer utilizing MRI? A recent meta-analysis by Dr. Li-Tao Zhao and colleagues at Beihang University in Beijing, China, aimed to answer this question.⁷ The authors wrote their paper because mortality secondary to prostate cancer is higher in China than in many western nations, and there is a higher proportion of patients with high-risk prostate cancer among the Chinese population due to limited availability of prostate-specific antigen screening.⁷ The authors noted that several AI software platforms have been approved for medical use by the US Food and Drug Administration in recent years, including ProstatID, software that helps radiologists identify suspicious areas on a prostate MRI. Furthermore, the integration of AI technologies within mpMRI clinical pathways may result in detection of higher rates of clinically significant prostate cancer, the authors wrote.

Existing literature has demonstrated the use of AI in various aspects of diagnosis and treatment of prostate cancer, including predictions of Gleason GG, biochemical recurrence, and extracapsular extension⁷; however, most review articles have focused on the process of creating AI models for these tasks, the authors noted. The number of review articles focusing on comparing results between clinical assessment methods and AI is limited.

The first segment of their review compared the 2 primary modalities for prostate MRI image segmentation. The authors explained that traditional AI models were developed using pretrained datasets of “handcrafted radiomics features” (eg, manually segmented areas of suspected lesions). These features are then transmitted to traditional machine learning models, such as support vector machines or ensemble techniques like random forest models, to help identify radiomic or clinical features of value.⁷ The issue with this type of AI model creation is that it is “time consuming, laborious, and can lead to subjective disagreements between radiologists,”⁷ the authors wrote. They reviewed an alternative method, which involves using “deep learning radiomics.” Essentially, it uses the concept of artificial neural networks to automatically extract important features from, in this case, prostate MRI features. The authors cited examples of existing models, such as ResNet,⁸ that have used deep learning to identify lesions of interest.

The second half of their review focused on evaluating and comparing AI-based techniques and clinical techniques for the diagnosis of prostate cancer. The authors found that deep learning-based AI techniques achieved an overall performance higher than any clinical assessments employed (area under the curve, 0.87 vs 0.82). Furthermore, benefits such as immediate “image to decision” diagnosis are simply not possible with clinical assessments that require manual image segmentation, the authors explained.

One of the primary benefits demonstrated in the meta-analysis is the combination of AI models and PI-RADS assessments, which can potentially decrease the number of prostate biopsies, resulting in decreased over-diagnosis and over-treatment of non-clinically significant prostate cancer while improving the specificity of PI-RADS.⁷

Despite the clear benefits of combining AI with existing clinical techniques, the authors cited several shortfalls with the existing literature, including, the use of small patient cohorts with external validation of models completed on retrospective datasets. AI research should focus on incorporating prospective data into the clinical pathway for AI-driven products focused on prostate cancer, they concluded.

Akhil Abraham Saji, MD, Fellow at the University of Southern California, is a urologist specializing in minimally invasive surgery and urologic oncology with an interest in technology-driven innovation within health care.

 

References

1. Cancer of the Prostate – Cancer Stat Facts. SEER. Accessed May 14, 2023. https://seer.cancer.gov/statfacts/html/prost.html
2. Wei JT, Barocas D, Carlsson S, et al. Early Detection of Prostate Cancer: AUA/SUO Guideline Part I: Prostate Cancer Screening. Journal of Urology. 2023;210(1):46-53. doi:10.1097/JU.0000000000003491
3. Murphy G, Haider M, Ghai S, Sreeharsha B. The Expanding Role of MRI in Prostate Cancer. American Journal of Roentgenology. 2013;201(6):1229-1238. doi:10.2214/AJR.12.10178
4. Eastham JA, Boorjian SA, Kirkby E. Clinically Localized Prostate Cancer: AUA/ASTRO Guideline. Journal of Urology. 2022;208(3):505-507. doi:10.1097/JU.0000000000002854
5. Hung AJ, Chen J, Ghodoussipour S, et al. A deep-learning model using automated performance metrics and clinical features to predict urinary continence recovery after robot-assisted radical prostatectomy. BJU international. 2019;124(3):487-495.
6. Chen J, Oh PJ, Cheng N, et al. Use of automated performance metrics to measure surgeon performance during robotic vesicourethral anastomosis and methodical development of a training tutorial. The Journal of urology. 2018;200(4):895-902.
7. Zhao LT, Liu ZY, Xie WF, et al. What benefit can be obtained from magnetic resonance imaging diagnosis with artificial intelligence in prostate cancer compared with clinical assessments? Military Med Res. 2023;10(1):29. doi:10.1186/s40779-023-00464-w
8. He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. ; 2016:770-778.